A069-07
Air Quality Modeling for Urban Environments Using Deep Neural Networks and Very High-Resolution Satellite Imagery
Wednesday, 9 December 2020: 05:54
Virtual
Michael von Pohle1, Adwait Sahasrabhojanee2, Ata Akbari Asanjan3, Emily Deardorff3, Violet Lingenfelter3, Meytar Sorek-Hamer4, Esra Suel5, Bryan Matthews6, Nikunj Oza7 and Michael Brauer8, (1)NASA Ames Research Center, USRA, Moffet Field, United States, (2)NASA Ames Research Center (USRA), Moffet Field, CA, United States, (3)NASA Ames Research Center (USRA), Moffet Field, United States, (4)NASA Ames Research Center, USRA, Moffett Field, United States, (5)Imperial College London, London, United Kingdom, (6)NASA Ames Research Center, SGT, Greenbelt, MD, United States, (7)NASA Ames Research Center, Mountain View, CA, United States, (8)University of British Columbia, School of Population and Public Health, Vancouver, BC, Canada
Abstract:
Air Quality (AQ) is currently measured using ground-based stations which require funding and infrastructure to support and are not spatially representative. In contrast, high-resolution commercial satellite imagery (⪯2 m/pixel) can be produced for almost any location on Earth and is readily available. The goal of this project is to develop a pipeline that uses MAXAR imagery to produce meter-scale, continuous maps of AQ for any city around the globe, with greater resolution and coverage than existing methods. To achieve this, we developed a Deep Neural Network (DNN) model that can produce AQ estimates for previously unseen urban locations using satellite imagery
Our DNN model is based on the VGG-16 neural network architecture, commonly used for object detection. We used this architecture to extract spatial features from satellite imagery. These features are passed on to Fully Connected (FC) layers to produce an estimate of PM2.5 and NO2 concentrations. We partitioned satellite imagery into 612,248 patches and fed these patches into the model to produce a 100 m (200 m) grid of PM2.5 (NO2) estimates over the study area. We trained the DNN model to predict modeled Land Use Regression (LUR) data over three cities (London, Vancouver, Los Angeles) and tested the model on a city (New York City) unseen by the model during training, achieving RMSE < 2 µg/m3 and a Pearson correlation of R=0.93 for all urban environments. Our results over New York City show potential for the model to be generalized for urban locations that don’t have modeled or measured AQ data.
Recent breakthroughs in DNNs can improve current capabilities in AQ estimation in the complete absence of ground measurements, and help address a global challenge in Earth Science.
